Green ammonia process real-time optimization method and system based on load prediction

Abstract

The present application relates to the technical field of green ammonia process optimization, process simulation, load prediction and artificial intelligence control. The method comprises: obtaining process structure information, equipment parameter information, historical operation data, renewable energy output data, storage unit state data, nitrogen supply state data and process constraint conditions of the green ammonia production process; constructing a green ammonia process mechanism process simulation model, and correcting the process simulation model based on the historical operation data; generating a future time domain load trajectory based on the renewable energy output prediction result, and constructing a training sample in combination with the process simulation result and the historical operation data; when it is found that the prediction deviation of the agent model exceeds a threshold value, extracting a failure sample and triggering directional retraining to realize online updating of the agent model. The present application can improve the flexible operation ability, energy utilization efficiency and green ammonia production stability of the system in view of the working condition uncertainty caused by the fluctuation of renewable energy such as wind and light.

Description

Technical Field

[0001] This invention relates to the fields of green ammonia process optimization, renewable energy coupling control, process simulation and artificial intelligence modeling, and specifically to a method and system for real-time optimization of green ammonia process parameters for renewable energy fluctuation scenarios. Background Technology

[0002] Green ammonia typically refers to a low-carbon ammonia production route that utilizes renewable energy sources such as wind and solar power to produce hydrogen through water electrolysis, nitrogen through air separation, and further ammonia synthesis. Compared to traditional fossil fuel-based ammonia synthesis processes, a significant characteristic of green ammonia systems is the greater volatility and uncertainty of their energy input. Affected by wind speed, irradiance, and grid fluctuations, the hydrogen production load and hydrogen supply rate of the electrolysis hydrogen production unit will change over time, thus affecting the hydrogen-nitrogen ratio, the inlet state of the synthesis tower, the cycle load, and the overall ammonia yield.

[0003] Existing optimization methods for green ammonia operation mainly fall into two categories: one is offline optimization methods based on mechanism process simulation. These methods can reflect the mechanistic laws of reaction, heat transfer, separation, and circulation processes. However, they require frequent recalculation under conditions of multiple time periods, dynamic fluctuations, and multivariate coupling, resulting in high computational costs and making it difficult to meet real-time optimization requirements. The other category is pure data-driven optimization methods. Although these methods have fast prediction speeds, they often ignore material conservation, energy conservation, equipment operating boundaries, and safety interlocking conditions. Under extrapolated operating conditions or drastic fluctuation conditions, they are prone to providing unexecutable parameter schemes.

[0004] Especially in scenarios involving fluctuations in renewable energy, green ammonia systems do not only need to handle a single steady-state condition, but also require real-time coordinated optimization of parameters such as synthesis pressure, circulation rate, cooling load, reactor inlet temperature, and hydrogen-nitrogen ratio under the combined effects of predicted load trajectories, changes in hydrogen buffer tank inventory, fluctuations in nitrogen supply, and changes in the thermal state of the synthesis loop. Therefore, there is an urgent need for a real-time optimization method and system for process parameters that can inherit the reliability of process simulation mechanisms, possess rapid predictive capabilities, and be specifically adapted to green ammonia fluctuation conditions. Summary of the Invention

[0005] I. Purpose of the Invention

[0006] The purpose of this invention is to address the problems of insufficient real-time optimization capability, poor reliability of pure data models, and slow process simulation solution speed in the existing green ammonia process under renewable energy fluctuation scenarios. It provides a method and system for real-time optimization of green ammonia process parameters for renewable energy fluctuation scenarios, so as to achieve rapid optimization of green ammonia process parameters under predicted load conditions, real-time correction under operating condition disturbances, and adaptive model update driven by failure samples.

[0007] II. Technical Solution

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A real-time optimization method for green ammonia processing based on load prediction, characterized by the following steps:

[0010] S1. Obtain process structure information, equipment parameter information, historical operating data, renewable energy output data, hydrogen storage status data, nitrogen supply status data, and process constraints of the green ammonia process.

[0011] S2. Based on the process structure information and equipment parameter information, establish a green ammonia process mechanism simulation model, and use the historical operating data to correct the parameters of the process simulation model;

[0012] S3. Perform load forecasting on the renewable energy output data, generate a predicted load trajectory within a preset time domain, and construct a time-varying operating condition input sequence based on the predicted load trajectory;

[0013] S4. Under the time-varying operating condition input sequence, sample the process parameters to be optimized, call the process simulation model to generate simulation sample data, and construct a training dataset by combining it with historical operating data.

[0014] S5. Train a machine learning agent model based on the training dataset, and introduce a process constraint penalty term into the loss function of the machine learning agent model to establish a mapping relationship between process parameters and key performance indicators of the process under load fluctuation conditions.

[0015] S6. Input the input conditions corresponding to the predicted load trajectory into the machine learning proxy model, and combine them with the intelligent optimization algorithm to generate candidate process parameter schemes;

[0016] S7. Input the candidate process parameter schemes into the process simulation model for simulation verification, and screen the target process parameter combinations that meet the process constraints.

[0017] S8. When the error between the process simulation verification result and the prediction result of the machine learning agent model exceeds the preset threshold, the corresponding candidate solution will be used as a failure sample to trigger the update of the machine learning agent model.

[0018] S9. Output the target process parameter combination and use it for real-time operation adjustment of the green ammonia process.

[0019] Preferably, in step S2, the parameter correction includes correcting one or more of the following: equivalent reaction kinetic parameters of the synthesis tower, overall heat transfer coefficient of the heat exchanger, compressor efficiency parameters, and separation unit efficiency parameters.

[0020] Preferably, in step S3, the input to the load forecast includes one or more of historical wind power, historical photovoltaic power, time characteristics, and meteorological characteristics, and the output is a sequence of predicted available power for multiple future control cycles.

[0021] Preferably, in step S4, the sampling method is stratified sampling, which divides the operating conditions into low load range, transitional load range and high load range according to the predicted load level, and sets different sampling densities in different ranges.

[0022] Preferably, in step S5, the process constraint penalty items include one or more of the following: material conservation deviation items, energy conservation deviation items, hydrogen-nitrogen ratio deviation items, equipment boundary violation penalty items, and safety interlock breach penalty items.

[0023] Preferably, the loss function of the machine learning agent model is: L = Lpred + αLmass + βLenergy + γLratio + δLbound + εLsafety, where Lpred is the prediction error term, Lmass is the material conservation deviation term, Lenergy is the energy conservation deviation term, Lratio is the hydrogen-nitrogen ratio deviation term, Lbound is the equipment boundary violation penalty term, Lsafety is the safety interlock breach penalty term, and α, β, γ, δ, and ε are weighting coefficients.

[0024] Preferably, in step S5, the machine learning agent model is a graph neural network model, a deep neural network model, a temporal prediction network model, or a combination of the above models.

[0025] Preferably, in step S6, the candidate process parameter scheme includes one or more of the following: synthesis tower inlet temperature, synthesis pressure, circulating gas flow rate, cooling load, hydrogen-nitrogen feed ratio, and venting air ratio.

[0026] This invention also provides a real-time optimization system for green ammonia process based on load forecasting, characterized by comprising: a data acquisition module, a mechanism modeling module, a load forecasting module, a sample construction module, a proxy modeling module, an optimization decision-making module, a simulation verification module, a failure sample analysis module, a model update module, and a control output module; wherein, the data acquisition module is used to acquire process structure information, equipment parameter information, historical operating data, renewable energy output data, hydrogen storage status data, nitrogen supply status data, and process constraints of the green ammonia process; the mechanism modeling module is used to establish a simulation model of the green ammonia process mechanism and perform parameter correction; the load forecasting module is used to predict the renewable energy output within a preset time domain and generate a predicted load trajectory. The sample construction module is used to construct a training dataset based on the predicted load trajectory and process simulation results; the surrogate modeling module is used to train a machine learning surrogate model containing process constraint penalty terms; the optimization decision module is used to generate candidate process parameter schemes by combining the predicted load trajectory and intelligent optimization algorithms; the simulation verification module is used to perform mechanistic simulation verification on the candidate process parameter schemes and screen the target process parameter combination; the failure sample analysis module is used to identify failure samples and extract their load fluctuation characteristics, constraint violation characteristics, and prediction error characteristics; the model update module is used to trigger local incremental updates or global retraining based on failure samples; and the control output module executes the method of sending the target process parameter combination to the green ammonia process control layer for execution.

[0027] III. Beneficial Effects

[0028] Compared with the prior art, the present invention has at least the following beneficial effects:

[0029] 1. This invention replaces the repetitive simulation calculations of large-scale green ammonia processes with machine learning proxy models, which can significantly improve the efficiency of process parameter search and optimization. It uses the load uncertainty caused by renewable energy fluctuations as the optimization trigger factor, and the parameter optimization results are more consistent with the characteristics of real chemical processes, thus improving the reliability of the results.

[0030] 2. This invention introduces a load prediction module, which expands the optimization object from a single steady-state condition to a rolling condition under the prediction time domain, thereby improving the real-time adjustment capability of the green ammonia system under wind and solar fluctuation scenarios.

[0031] 3. This invention does not simply adopt a general machine learning model, but embeds material conservation, energy conservation, hydrogen-nitrogen ratio constraints, equipment boundaries, and safety interlocking conditions into the loss function, thereby improving the feasibility judgment ability of the surrogate model in chemical process scenarios;

[0032] 4. This invention achieves closed-loop collaboration between the mechanistic model and the proxy model through process simulation verification and failure sample classification and update mechanism, thereby improving the model's generalization ability and long-term stability under dynamic disturbance conditions.

[0033] 5. This invention can achieve rapid real-time optimization of green ammonia process parameters while ensuring ammonia yield, equipment safety and system stability, thereby improving energy utilization, reducing wind and solar curtailment rates and enhancing the flexibility of green ammonia production. Attached Figure Description

[0034] Figure 1 This is a flowchart of a method for real-time optimization of green ammonia process parameters for renewable energy fluctuation scenarios provided in an embodiment of the present invention;

[0035] Figure 2 A block diagram of the real-time optimization system for green ammonia process parameters provided in this embodiment of the invention;

[0036] Figure 3 This is a schematic diagram illustrating the collaborative optimization of load forecasting, process simulation, and proxy model provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram of closed-loop optimization for triggering updates based on failed samples, provided in an embodiment of the present invention. Detailed Implementation

[0038] The inventive concept of this invention is:

[0039] Traditional green ammonia systems are highly susceptible to fluctuations in renewable energy sources such as wind and solar power, with hydrogen production loads constantly changing. While existing optimization methods, including mechanistic process simulation, can reflect physicochemical laws, their high computational cost makes them unsuitable for real-time optimization. Pure data-driven models, though fast in prediction, often neglect material conservation, energy conservation, and equipment safety boundaries, potentially leading to unenforceable and dangerous parameters during periods of drastic operational fluctuations. This invention aims to overcome this contradiction between computational speed and physical reliability. Core Solution: The core concept of this invention lies in constructing a "machine learning proxy model embedded with physical mechanisms" and designing a closed-loop collaborative mechanism of "prediction-optimization-verification-update." This proxy model replaces traditional large-scale repetitive simulation calculations to accelerate the process, while rigorous verification and adaptive model evolution through process simulation ensure the absolute safety and feasibility of the output parameters. Key Technological Innovations: Proactive Rolling Operating Condition Construction: Breaking through the limitations of traditional single steady-state optimization, this invention introduces short-term load forecasting, expanding the optimization object to a time-varying operating condition input sequence in the prediction time domain, proactively adapting to the fluctuations in renewable energy. Reconstruction of the loss function based on physical mechanism constraints: Instead of using a general black-box model, this invention innovatively embeds material conservation deviation, energy conservation deviation, hydrogen-nitrogen ratio deviation, equipment exceeding limits, and safety interlock violation penalties directly into the loss function of the machine learning proxy model. This endows the model with the ability to judge physical feasibility in chemical engineering scenarios. Failure sample-driven closed-loop model evolution: Candidate solutions generated by the proxy model are input into the process simulation model for mechanistic simulation verification. When the error between the two exceeds a preset threshold, "failure samples" are accurately captured, and local incremental updates or global retraining of the proxy model are triggered based on the operating condition range and violation type. Expected technical effects: This invention achieves complementary advantages between the mechanistic model and the proxy model, enabling millisecond-level rapid real-time optimization of green ammonia process parameters while ensuring ammonia yield, equipment safety, and system stability. This significantly improves the flexible operation and real-time adjustment capabilities of the green ammonia system under fluctuating wind and solar power conditions.

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0041] Example 1: Real-time optimization method for green ammonia process parameters in scenarios with renewable energy fluctuations

[0042] like Figure 1 As shown in the figure, this embodiment provides a method for real-time optimization of green ammonia process parameters for renewable energy fluctuation scenarios, including the following steps.

[0043] Step S1: Data Acquisition and Preprocessing

[0044] The basic data of the green ammonia process is obtained, including: (1) process structure information, including the electrolysis hydrogen production unit, hydrogen buffer unit, air separation nitrogen production unit, ammonia synthesis loop, heat exchange unit, compression unit and their connection relationship; (2) equipment parameter information, including the rated power of the electrolyzer, the volume of the hydrogen buffer tank, the parameters of the synthesis tower bed, the compressor efficiency, the heat transfer area of ​​the heat exchanger, etc.; (3) historical operating data, including wind and solar power output, hydrogen production load, hydrogen storage inventory, temperature, pressure, flow rate, ammonia yield and energy consumption, etc.; (4) process constraints, including the upper and lower limits of equipment operation, product quality constraints, safety interlock constraints and hydrogen storage inventory constraints. The historical operating data is processed by missing value completion, outlier removal, time sequence alignment and normalization.

[0045] This step comprehensively acquires multi-source heterogeneous data covering process structure, equipment parameters, historical operation, energy output, and process constraints, laying a solid data foundation for subsequent mechanistic modeling and surrogate model training. Preprocessing operations such as missing value completion and outlier removal are performed on the historical operation data, effectively eliminating the interference of on-site measurement noise on model accuracy and improving the overall data quality and reliability of the optimization system.

[0046] Step S2: Construction and calibration of process simulation model

[0047] Based on the process structure and equipment parameter information obtained in step S1, a process simulation model of the green ammonia process mechanism is established in the process simulation platform. The process simulation model covers at least the electrolytic hydrogen production unit, the hydrogen-nitrogen ratio unit, the compression and heat exchange unit, the synthesis tower, and the separation loop. Historical operating data is used to calibrate the parameters of the process simulation model, including correcting the equivalent reaction kinetic parameters of the synthesis tower by minimizing the deviation between the simulation output and the field measurements, correcting the overall heat transfer coefficient of the heat exchanger using heat exchange data at historical load points, and correcting the compressor efficiency parameters based on historical operating power and outlet pressure data.

[0048] This step established a simulation model of the green ammonia process that strictly follows the physicochemical mechanisms, and used historical data to correct the synthesis tower kinetics, heat exchanger heat transfer coefficients, and compressor efficiency. This step effectively reduced the deviation between theoretical simulation and actual plant operating conditions, ensuring the absolute reliability of the process simulation model in subsequent sample generation and as a "verification gatekeeper."

[0049] Step S3: Load Forecasting and Sample Construction

[0050] Short-term load forecasting was performed on renewable energy output data to obtain a predicted available power sequence for multiple future control cycles. A time-varying operating condition input sequence was constructed by combining hydrogen storage tank inventory status and nitrogen supply capacity. The operating conditions were divided into low-load, transitional-load, and high-load intervals based on the predicted load level, and a stratified sampling method was used to sample synthesis pressure, circulation flow rate, cooling load, hydrogen-nitrogen ratio, and venting ratio. The sampling results were input into the process simulation model to obtain corresponding key performance indicators such as ammonia yield, energy consumption, compressor power, and synthesis tower outlet temperature. A training dataset was constructed by fusing the above simulation samples with historical operating data.

[0051] This step introduces a short-term load forecasting mechanism, transforming the system's control strategy from "passive response" to "active forecasting," significantly enhancing the green ammonia system's forward-looking adjustment capability in response to wind and solar load fluctuations. The adoption of a tiered sampling strategy (using different sampling densities for low, transitional, and high load ranges) effectively overcomes the uneven data distribution problem caused by renewable energy fluctuations, significantly improving the surrogate model's sample coverage and prediction accuracy under boundary conditions such as severe fluctuations.

[0052] Step S4: Proxy Model Training

[0053] The machine learning proxy model is trained based on the training dataset constructed in step S3. The inputs to the proxy model include predicted load characteristics, hydrogen storage state characteristics, nitrogen supply characteristics, and candidate process parameters. The outputs include ammonia yield, unit ammonia energy consumption, synthesis tower outlet temperature, compressor power, and constraint feasibility indicators. To enhance the model's process adaptability, material conservation deviation terms, energy conservation deviation terms, hydrogen-nitrogen ratio deviation terms, equipment boundary violation penalties, and safety interlock breach penalties are introduced into the loss function.

[0054] This step is one of the core innovations of this invention. By directly embedding penalty terms such as material conservation, energy conservation, hydrogen-nitrogen ratio, and safety limit violations into the loss function of the machine learning proxy model, the model is forced to learn and comply with the laws of chemical physics during the optimization process. This completely solves the fatal flaw of traditional pure data-driven models, which are prone to outputting dangerous process parameters that "defy common sense" or "cause safety accidents" under unknown or drastically fluctuating operating conditions.

[0055] Step S5: Fast Optimization Search Based on Proxy Model

[0056] The current time or rolling time domain corresponding to the predicted load trajectory is input into a machine learning surrogate model, and candidate process parameter schemes are generated by combining them with intelligent optimization algorithms. The optimization objective can be set as minimizing unit ammonia energy consumption under the premise of meeting the lower limit of ammonia yield and safety constraints, or maximizing the cumulative ammonia production in the future time domain under hydrogen storage inventory constraints. Intelligent optimization algorithms can adopt genetic algorithms, particle swarm optimization algorithms, or Bayesian optimization algorithms.

[0057] This step utilizes a pre-trained surrogate model to replace the computationally intensive traditional mechanism simulation for iterative optimization of massive solutions. This significantly reduces computational complexity, compressing optimization time from "minutes / hours" to "milliseconds / seconds," truly meeting the real-time optimization needs of renewable energy high-frequency fluctuation scenarios.

[0058] Step S6: Simulation verification of candidate solutions

[0059] The candidate solutions generated in step S5 are re-input into the process simulation model for simulation verification to determine whether they meet the upper limit of synthesis tower temperature, upper limit of system pressure, upper limit of compressor power, feasible range of hydrogen-nitrogen ratio, lower limit of hydrogen storage safety, and safety interlock requirements. If all constraints are met, the solution is retained as a feasible solution; otherwise, it is marked as an infeasible solution.

[0060] This step involves re-inputting the fast optimal solution provided by the surrogate model into the high-precision mechanistic model for rigorous verification. This is equivalent to adding a "double safety insurance" to the real-time control system, 100% intercepting any out-of-limit solutions that may arise due to the failure of the surrogate model's extrapolation, ensuring that the process parameters finally issued to the DCS / APC control layer are absolutely safe and feasible.

[0061] Step S7: Failed samples trigger update and result output

[0062] When the error between the surrogate model's prediction and the process simulation verification results exceeds a preset threshold, the corresponding sample is extracted as a failure sample. This failure sample is then categorized based on its load fluctuation range, error type, and default type, triggering either a local incremental update or a global retraining. Finally, the target process parameter combination is output and sent to the upper-level scheduling system or control layer for execution. This mechanism forms a closed-loop optimization process of "load prediction - simulation - optimization - verification - update".

[0063] This step establishes a closed-loop adaptive evolution mechanism. When the surrogate model fails to predict, the erroneous sample is not discarded but instead extracted as a high-value "adversarial sample," triggering local or global model updates. This enables the optimization system to "learn and correct itself," adapting to long-term dynamic disturbances such as catalyst deactivation, equipment aging, or seasonal weather changes, ensuring long-term high accuracy and stability of the model throughout its lifecycle.

[0064] The following section explains the synergistic technical effects (overall technical effects) of the combination of each step.

[0065] By treating steps S1 to S7 as an organic whole, a system-level combined technical effect of "1+1>2" can be achieved: a perfect unity of "reliable mechanism" and "efficient data" (S2 + S4 + S5): Traditional technologies often compromise between "accurate calculation but too slow" (mechanism simulation) and "fast calculation but unsafe" (pure data model). The combined steps of this invention use a proxy model with embedded mechanism constraints (S4) for fast search (S5), and use the corrected mechanism model (S2) as a benchmark, successfully breaking this contradiction and achieving real-time optimization that balances high computational speed and high physical reliability. Flexible closed-loop control for "proactively defending against fluctuations" (S3 + S6 + S7): Faced with random fluctuations in wind and solar power, this invention not only achieves preemptive parameter adjustment preparation through S3's advance prediction of load trajectory, but also constructs a closed-loop system of "prediction-verification-failure feedback-model evolution" through S6 and S7. This combined effect enables green ammonia production units to function like intelligent and resilient living organisms, maximizing the absorption of fluctuating renewable energy sources while ensuring safe operation, significantly reducing wind and solar curtailment rates and improving system flexibility.

[0066] Example 2: Real-time optimization system for green ammonia process parameters

[0067] like Figure 2 As shown, this embodiment provides a real-time optimization system for green ammonia process parameters, including:

[0068] The system comprises a data acquisition module, a mechanism modeling module, a load forecasting module, a sample construction module, a surrogate modeling module, an optimization decision-making module, a simulation verification module, a failure sample analysis module, a model update module, and a control output module. Furthermore, the optimization decision-making module may also include a rolling optimization unit and a scheme ranking unit. The rolling optimization unit performs real-time optimization for multiple future control steps in the rolling time domain, while the scheme ranking unit prioritizes candidate schemes that meet the constraints based on their comprehensive objective function values.

[0069] Example 3: Application Scenarios of Green Ammonia

[0070] Taking the wind-solar coupled electrolysis hydrogen production-ammonia synthesis process as an example, the target system includes a wind power unit, a photovoltaic unit, a water electrolysis hydrogen production unit, a hydrogen buffer tank, an air separation nitrogen production unit, a compression and heat exchange unit, and an ammonia synthesis loop. The synthesis pressure, synthesis tower inlet temperature, circulating gas flow rate, hydrogen-nitrogen feed ratio, and venting ratio are used as process parameters to be optimized; unit ammonia energy consumption, instantaneous ammonia yield, and compressor power are used as key performance indicators; and the lower limit of hydrogen storage, the upper limit of synthesis tower outlet temperature, the upper limit of synthesis system pressure, and safety interlock requirements are used as process constraints. By adopting the method of this invention, process parameters can be adjusted in advance based on the predicted available power sequence for multiple future control cycles. When renewable energy output suddenly drops, priority is given to maintaining system safety and hydrogen balance; when renewable energy output increases, the ammonia synthesis load is increased, thereby enhancing the flexible operation capability of the green ammonia system.

[0071] Example 4: Electronic Device Example

[0072] This embodiment provides an electronic device, including a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the real-time optimization method for green ammonia process parameters in renewable energy fluctuation scenarios described in Embodiment 1 above. The electronic device can be a server, an industrial control terminal, an engineer's workstation, or a cloud computing node.

[0073] Example 5: Example of a computer-readable storage medium

[0074] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the method described in Embodiment 1 above. The computer-readable storage medium may be a hard disk, a solid-state drive, flash memory, an optical disk, or other media capable of storing program code.

[0075] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time optimization method for ammonia greening process based on load prediction, characterized in that, Includes the following steps: S1. Obtain process structure information, equipment parameter information, historical operating data, renewable energy output data, hydrogen storage status data, nitrogen supply status data, and process constraints of the green ammonia process. S2. Based on the process structure information and equipment parameter information, establish a green ammonia process mechanism simulation model, and use the historical operating data to correct the parameters of the process simulation model; S3. Perform load forecasting on the renewable energy output data, generate a predicted load trajectory within a preset time domain, and construct a time-varying operating condition input sequence based on the predicted load trajectory; S4. Under the time-varying operating condition input sequence, sample the process parameters to be optimized, call the process simulation model to generate simulation sample data, and construct a training dataset by combining it with historical operating data. S5. Train a machine learning agent model based on the training dataset, and introduce a process constraint penalty term into the loss function of the machine learning agent model to establish a mapping relationship between process parameters and key performance indicators of the process under load fluctuation conditions. S6. Input the input conditions corresponding to the predicted load trajectory into the machine learning proxy model, and combine them with the intelligent optimization algorithm to generate candidate process parameter schemes; S7. Input the candidate process parameter schemes into the process simulation model for simulation verification, and screen the target process parameter combinations that meet the process constraints. S8. When the error between the process simulation verification result and the prediction result of the machine learning agent model exceeds the preset threshold, the corresponding candidate solution will be used as a failure sample to trigger the update of the machine learning agent model. S9. Output the target process parameter combination and use it for real-time operation adjustment of the green ammonia process.

2. The method according to claim 1, characterized in that, In step S2, the parameter correction includes correcting one or more of the following: equivalent reaction kinetic parameters of the synthesis tower, overall heat transfer coefficient of the heat exchanger, compressor efficiency parameters, and separation unit efficiency parameters.

3. The method according to claim 1, characterized in that, In step S3, the inputs to the load forecast include one or more of historical wind power, historical photovoltaic power, time characteristics, and meteorological characteristics, and the output is a sequence of predicted available power for multiple future control cycles.

4. The method according to claim 1, characterized in that, In step S4, the sampling method is stratified sampling, which divides the operating conditions into low load range, transitional load range and high load range according to the predicted load level, and sets different sampling densities in different ranges.

5. The method according to claim 1, characterized in that, In step S5, the process constraint penalty items include one or more of the following: material conservation deviation items, energy conservation deviation items, hydrogen-nitrogen ratio deviation items, equipment boundary violation penalty items, and safety interlock breach penalty items.

6. The method according to claim 5, characterized in that, The loss function of the machine learning agent model is: L = L pred +αL mass +βL energy +γL ratio +δL bound +εL safety , where L pred For the prediction error term, L mass L is the material conservation deviation term. energy L is the energy conservation deviation term. ratio For the hydrogen-nitrogen ratio deviation term, L bound L is a penalty item for exceeding equipment boundary limits. safety For the penalty item for breach of safety interlock, α, β, γ, δ, and ε are weighting coefficients.

7. The method according to claim 1, characterized in that, In step S5, the machine learning agent model is a graph neural network model, a deep neural network model, a temporal prediction network model, or a combination of the above models.

8. The method according to claim 1, characterized in that, In step S6, the candidate process parameter schemes include one or more of the following: synthesis tower inlet temperature, synthesis pressure, circulating gas flow rate, cooling load, hydrogen-nitrogen feed ratio, and venting air ratio.

9. A real-time optimization system for green ammonia process based on load prediction, characterized in that, include: The system comprises a data acquisition module, a mechanism modeling module, a load forecasting module, a sample construction module, a surrogate modeling module, an optimization decision-making module, a simulation verification module, a failure sample analysis module, a model update module, and a control output module. The data acquisition module acquires process structure information, equipment parameter information, historical operating data, renewable energy output data, hydrogen storage status data, nitrogen supply status data, and process constraints for the green ammonia process. The mechanism modeling module establishes a simulation model of the green ammonia process mechanism and performs parameter correction. The load forecasting module predicts renewable energy output within a preset time domain and generates a predicted load trajectory. The sample construction module constructs a training dataset based on the predicted load trajectory and process simulation results. The surrogate modeling module trains a machine learning surrogate model that includes process constraint penalties. The optimization decision-making module combines the predicted load trajectory and intelligent optimization algorithms to generate candidate process parameter schemes. The simulation verification module is used to perform mechanism simulation verification on candidate process parameter schemes and screen target process parameter combinations; the failure sample analysis module is used to identify failure samples and extract their load fluctuation characteristics, constraint default characteristics and prediction error characteristics; the model update module is used to trigger local incremental updates or global retraining based on failure samples; the control output module executes the method of sending the target process parameter combination to the green ammonia process control layer for execution.

Images